Quick answer
Deploy an AI SDR in stages. Send from separate domains, never your main one, and warm them up for 3 to 4 weeks. Start with a person approving every message, and loosen that only after a few weeks with no complaints and almost no edits. Keep volume at 20 to 30 emails per mailbox per day and the spam rate in Google Postmaster Tools under 0.1%. Give the agent a verified list, an exclusion list and a one-page rules sheet for tone and topics. Keep a person on objections, pricing, angry replies, customers, open deals and data requests. Automate the work, not the thinking.
Before you switch it on
This is about rolling out an AI SDR you've already chosen. If you're still choosing, read what AI SDRs do well and where they fall short and how to vet an AI SDR vendor's claims first.
I set up cold email and LinkedIn outbound with B2B teams, and with an AI SDR the rollout decides more than the tool does. Have five things ready before the agent writes a word:
- An offer that has booked meetings when a person sent it. An agent scales what works. It won't find it for you.
- A written definition of who to contact: industries, company size, roles and the signals that make someone worth contacting now.
- One named owner who approves messages and answers the replies the agent hands over.
- A one-page rules sheet, covered in step 4.
- A contract you can leave, with an answer in writing on who keeps the domains, mailboxes and data when you do.
Step 1: send from separate domains, not your main one
Your main domain carries your invoices, your support email and every message your team sends. If an AI SDR gets it flagged as spam, all of that suffers. So the agent sends from separate domains that look like yours, such as getacme.com next to acme.com, and point visitors to your site.
- Authenticate each one. Google's sender guidelines ask you to set up authentication for each of your sending domains: SPF, DKIM and DMARC.
- Warm them up. New domains and mailboxes need 3 to 4 weeks of warmup before real volume. The routine is in how to warm up cold email domains.
- Ask whose they are. Some plans include them. 11x's pricing page lists managed Gmail mailboxes with domain setup and warmup, and AiSDR's plans come with a set number of domains and mailboxes. Find out whether you keep them if you cancel. If the domains leave with the vendor, so does their sending history.
Step 2: start with a person approving every message
Most AI SDRs can draft and wait for a yes. Salesforge lists "Auto-Pilot & Co-Pilot Modes" for its Agent Frank, and Artisan says that with Ava you set the guardrails yourself: approvals, escalation rules and when a human steps in. (I've written guest articles for Salesforge's blog. No tool in this article pays me.) Ask for this mode in the demo and use it. Then move through four stages, and only when the exit rule is met:
| Stage | The agent | A person | Move on when |
|---|---|---|---|
| 1. Shadow, during warmup | Researches and drafts, sends nothing | Reads every draft and fixes the rules sheet | Most drafts need no edits |
| 2. Approve everything | Sends only what was approved | Approves every email and LinkedIn message | 2 to 3 weeks with almost no edits and no complaints |
| 3. Approve first touches | Sends follow-ups on its own | Approves first emails, reads a sample of follow-ups, handles every hand-over | 4 more clean weeks, spam rate under 0.1% |
| 4. Rules only | Sends within the rules sheet | Reads 20 conversations a week, handles hand-overs | Stays here while the weekly checks hold |
Go back a stage whenever something slips: a complaint, a wrong fact about a prospect, a reply that got the wrong answer. The temptation is to switch approvals off in the quiet weeks, when the drafts have read fine for a while. The review is there for the one send you didn't expect.
What a missing review step costs: February 2026
On 12 February 2026, Prolific North reported on a cold email with the subject line "Saw your name in the Epstein Files", sent by Clustox, a US software and technology consultancy. A Manchester agency founder who got it posted it on LinkedIn and the post spread. Commenters said they had received similar emails in the days before. Clustox apologised publicly. It said a member of its outreach team had fallen "significantly short" of its standards and that it was overhauling its outreach review process so every message gets vetted for tone.
No report says AI wrote that line, and I'm not claiming it did. The lesson holds either way. The fix Clustox announced afterwards, a review step for tone, is stage 2 of this rollout, so put it in before launch. And an agent sending on its own can repeat one bad idea across a whole segment before anyone reads a reply. Decide two things on day one: who can pause every sequence within minutes and who answers in public if something goes out that shouldn't.
Step 3: keep the volume low and steady
Speed is the point of an AI SDR, and it's also how it hurts you fastest. Three rules:
- 20 to 30 emails per mailbox per day. Add mailboxes to grow, not emails per mailbox.
- Ramp slowly. Google's guidelines say to start with a low volume to engaged users, send at a consistent rate and avoid sudden spikes.
- Watch the spam rate. Google asks senders to keep the spam rate in Postmaster Tools below 0.1% and never reach 0.3%.
Work out the mailboxes before you pick a plan. 1,000 new contacts a month with 3 emails each is 3,000 emails. At 25 a day over about 21 sending days, one mailbox sends about 525 a month, so you need 6 mailboxes. If a plan includes fewer for its contact allowance, each mailbox gets pushed harder than it should.
Step 4: give it clean data and a one-page rules sheet
An agent with a bad list fails faster and at higher volume than a person would. Before launch:
- Verify every email address, so bounces don't hurt the new domains.
- Load an exclusion list: customers, open deals, partners, competitors, past complaints and everyone who unsubscribed. Sync it with your CRM so a new deal drops out of the sequence the day it opens.
- Hand over only the data it needs, and read what the contract says about training on your data and getting it back.
Then write the rules sheet. One page, readable in two minutes:
- Tone, with three openers you like and three you'd never send.
- Banned topics: news, tragedy, politics, health, jokes at the prospect's expense, fake urgency and pretending to know someone.
- Allowed facts: which customer names, results and prices the agent may mention. Everything else stays out.
- Hand-over triggers: the replies that go straight to a person, listed in step 5.
Here's the kind of draft the review step should catch:
"Hi Sarah, congrats on the amazing Series B! Loved your recent post. Quick question: are you still doing outbound manually?"
And what should go out instead, one real reason and one question:
"Hi Sarah, you just opened a Berlin office. Curious how you're handling outbound for the region?"
The first one flatters and refers to a post the agent may never have read. The second names something that happened and asks one thing.
Step 5: decide what stays human
Automate the work, not the thinking. AI owns the speed, you own the judgment. The agent can research, draft, send and follow up. These go to a person every time:
| Situation | Why a person handles it |
|---|---|
| Objections and pricing questions | A wrong answer loses the deal or promises something you can't deliver |
| Angry replies, threats to post publicly, "is this a bot?" | Another automated touch turns one annoyed person into a screenshot |
| Existing customers, open deals, partners | The context lives outside the sequence |
| Senior people at your top accounts | One clumsy message can close the door for a year |
| Requests to see or delete their data | It's a legal request, so log it and answer it properly |
| Anything touching news, someone's personal life or a sensitive topic | Getting the tone wrong happens in public |
Pause the sequence for that contact the moment a reply is handed over, so the next scheduled follow-up doesn't land on top of a complaint.
A quick test. Read the last ten replies the agent answered on its own. If you'd have written even one of them differently in a way that matters, it's not ready for the next stage.
Step 6: go slower on LinkedIn than on email
When I set up outbound for clients, email and LinkedIn run in one sequence. An AI SDR that also sends on LinkedIn carries a different risk, because what's at stake is a person's own profile, not a spare domain. LinkedIn's User Agreement (section 8.2) says you may not use "bots or other unauthorized automated methods" to add contacts or send messages.
- Keep LinkedIn steps in the approval stage longer than email.
- Keep daily invites and messages well below what the tool allows.
- Don't start on the founder's or CEO's profile.
The limits and the risk in more detail: is LinkedIn automation safe.
Step 7: check it every week
- Spam rate in Postmaster Tools, under 0.1%.
- Bounces and unsubscribes, against the week before.
- 20 conversations, read start to finish. Look for wrong facts, odd tone and the same opener repeated across prospects.
- Hand-overs: did each one reach a person the same day?
- Held meetings and what the account executive thought of them. Meetings that go nowhere point to a targeting problem.
If two of these slip in the same week, drop back a stage in step 2 until they recover.
Key takeaways
- Send from separate, authenticated domains warmed for 3 to 4 weeks, and find out who keeps them if you cancel.
- Move through approval stages: shadow, approve everything, approve first touches, rules only. Go back a stage when something slips.
- Keep volume at 20 to 30 emails per mailbox per day and the Google spam rate under 0.1%.
- Feed it a verified list, an exclusion list synced with your CRM and a one-page rules sheet.
- Objections, pricing, angry replies, customers, open deals, top accounts and data requests go to a person.
FAQ
How do you deploy an AI SDR safely?
In stages: separate sending domains warmed for 3 to 4 weeks, a person approving every message at first, low and steady volume, a verified list with exclusions and a one-page rules sheet. Loosen approvals only after weeks without complaints or edits.
Should an AI SDR send from my main domain?
No. Send from separate domains that point to your site, each with SPF, DKIM and DMARC set up. If a sending domain gets flagged, your main domain, with your invoices and support email, stays clean.
How long should a person approve every AI SDR message?
Until the drafts need almost no edits and nothing has gone wrong for 2 to 3 weeks after warmup. Then approve first emails only and read a weekly sample. Go back to full approval whenever something slips.
How many emails a day should an AI SDR send?
20 to 30 per mailbox per day, with more mailboxes when you need more volume. Ramp up slowly, as Google's sender guidelines advise, and keep the spam rate in Postmaster Tools under 0.1%.
What should an AI SDR never handle on its own?
Objections, pricing questions, angry replies, existing customers, open deals, senior people at your top accounts and requests to see or delete personal data. Hand those to a person and pause the sequence for that contact.